Relation extraction for the impatient knowledge builder

What it does OpenNRE is a toolkit for neural relation extraction. Given a sentence and the positions of two entities, it predicts the relation between them—turning Bill Gates founded Microsoft into a structured triple. It supports both sentence-level supervised learning and bag-level distant supervision, aggregating multiple sentences that mention the same entity pair.

The interesting bit The toolkit abstracts away model-specific plumbing, offering a single interface for both CNN-based models (with attention) and BERT. It also ships with several pretrained checkpoints trained on Wiki80 and TACRED, so you can load a model and call infer() immediately without training from scratch.

Key highlights

• Supports both CNN-based encoders (including CNN-ATT) and pretrained language models like BERT under one interface
• Handles both supervised sentence-level and distantly supervised bag-level relation extraction
• Ships with ready-to-use pretrained models for Wiki80 and TACRED datasets
• Evaluation on NYT10 shows its CNN-ATT implementation slightly outperforms the original paper (AUC 0.333 vs. 0.318)
• Modular design aimed at researchers who want to swap model components with minimal code changes

Caveats

• Still source-only; no PyPI package is available yet
• Datasets and pretrained weights are not bundled and must be fetched separately via provided scripts
• Requires manual matching of PyTorch and CUDA versions

Verdict Worth exploring if you are building knowledge graphs or need a modular research sandbox for relation extraction. If you want a polished, pip-installable library with bundled weights, look elsewhere for now.